*
* Scheduling and dispatching. */
-#ifdef __SHARC__
-#pragma nosharc
-#endif
-
#include <schedule.h>
+#include <corerequest.h>
#include <process.h>
#include <monitor.h>
#include <stdio.h>
#include <manager.h>
#include <alarm.h>
#include <sys/queue.h>
+#include <arsc_server.h>
+#include <hashtable.h>
/* Process Lists. 'unrunnable' is a holding list for SCPs that are running or
* waiting or otherwise not considered for sched decisions. */
struct proc_list unrunnable_scps = TAILQ_HEAD_INITIALIZER(unrunnable_scps);
struct proc_list runnable_scps = TAILQ_HEAD_INITIALIZER(runnable_scps);
-struct proc_list all_mcps = TAILQ_HEAD_INITIALIZER(all_mcps);
-spinlock_t sched_lock = SPINLOCK_INITIALIZER;
-
-// This could be useful for making scheduling decisions.
-/* Physical coremap: each index is a physical core id, with a proc ptr for
- * whoever *should be or is* running. Very similar to current, which is what
- * process is *really* running there. */
-struct proc *pcoremap[MAX_NUM_CPUS];
-
-/* Tracks which cores are idle, similar to the vcoremap. Each value is the
- * physical coreid of an unallocated core. */
-spinlock_t idle_lock = SPINLOCK_INITIALIZER;
-uint32_t idlecoremap[MAX_NUM_CPUS];
-uint32_t num_idlecores = 0;
-uint32_t num_mgmtcores = 1;
+/* mcp lists. we actually could get by with one list and a TAILQ_CONCAT, but
+ * I'm expecting to want the flexibility of the pointers later. */
+struct proc_list all_mcps_1 = TAILQ_HEAD_INITIALIZER(all_mcps_1);
+struct proc_list all_mcps_2 = TAILQ_HEAD_INITIALIZER(all_mcps_2);
+struct proc_list *primary_mcps = &all_mcps_1;
+struct proc_list *secondary_mcps = &all_mcps_2;
/* Helper, defined below */
-static void __core_request(struct proc *p);
-static void __put_idle_cores(uint32_t *pc_arr, uint32_t num);
+static void __core_request(struct proc *p, uint32_t amt_needed);
static void add_to_list(struct proc *p, struct proc_list *list);
static void remove_from_list(struct proc *p, struct proc_list *list);
static void switch_lists(struct proc *p, struct proc_list *old,
struct proc_list *new);
+static void __run_mcp_ksched(void *arg); /* don't call directly */
+static uint32_t get_cores_needed(struct proc *p);
+
+/* Locks / sync tools */
+
+/* poke-style ksched - ensures the MCP ksched only runs once at a time. since
+ * only one mcp ksched runs at a time, while this is set, the ksched knows no
+ * cores are being allocated by other code (though they could be dealloc, due to
+ * yield).
+ *
+ * The main value to this sync method is to make the 'make sure the ksched runs
+ * only once at a time and that it actually runs' invariant/desire wait-free, so
+ * that it can be called anywhere (deep event code, etc).
+ *
+ * As the ksched gets smarter, we'll probably embedd this poker in a bigger
+ * struct that can handle the posting of different types of work. */
+struct poke_tracker ksched_poker = POKE_INITIALIZER(__run_mcp_ksched);
+
+/* this 'big ksched lock' protects a bunch of things, which i may make fine
+ * grained: */
+/* - protects the integrity of proc tailqs/structures, as well as the membership
+ * of a proc on those lists. proc lifetime within the ksched but outside this
+ * lock is protected by the proc kref. */
+//spinlock_t proclist_lock = SPINLOCK_INITIALIZER; /* subsumed by bksl */
+/* - protects the provisioning assignment, and the integrity of all prov
+ * lists (the lists of each proc). */
+//spinlock_t prov_lock = SPINLOCK_INITIALIZER;
+/* - protects allocation structures */
+//spinlock_t alloc_lock = SPINLOCK_INITIALIZER;
+spinlock_t sched_lock = SPINLOCK_INITIALIZER;
/* Alarm struct, for our example 'timer tick' */
struct alarm_waiter ksched_waiter;
set_alarm(&per_cpu_info[core_id()].tchain, &ksched_waiter);
}
-/* Kmsg, to run the scheduler tick (not in interrupt context) and reset the
+/* RKM alarm, to run the scheduler tick (not in interrupt context) and reset the
* alarm. Note that interrupts will be disabled, but this is not the same as
* interrupt context. We're a routine kmsg, which means the core is in a
* quiescent state. */
-static void __ksched_tick(struct trapframe *tf, uint32_t srcid, long a0,
- long a1, long a2)
+static void __ksched_tick(struct alarm_waiter *waiter)
{
/* TODO: imagine doing some accounting here */
- schedule();
- /* Set our alarm to go off, incrementing from our last tick (instead of
- * setting it relative to now, since some time has passed since the alarm
- * first went off. Note, this may be now or in the past! */
- set_awaiter_inc(&ksched_waiter, TIMER_TICK_USEC);
+ run_scheduler();
+ /* Set our alarm to go off, relative to now. This means we might lag a bit,
+ * and our ticks won't match wall clock time. But if we do incremental,
+ * we'll actually punish the next process because the kernel took too long
+ * for the previous process. Ultimately, if we really care, we should
+ * account for the actual time used. */
+ set_awaiter_rel(&ksched_waiter, TIMER_TICK_USEC);
set_alarm(&per_cpu_info[core_id()].tchain, &ksched_waiter);
}
-/* Interrupt/alarm handler: tells our core to run the scheduler (out of
- * interrupt context). */
-static void __kalarm(struct alarm_waiter *waiter)
-{
- send_kernel_message(core_id(), __ksched_tick, 0, 0, 0, KMSG_ROUTINE);
-}
-
void schedule_init(void)
{
- TAILQ_INIT(&runnable_scps);
- TAILQ_INIT(&all_mcps);
+ spin_lock(&sched_lock);
assert(!core_id()); /* want the alarm on core0 for now */
- init_awaiter(&ksched_waiter, __kalarm);
+ init_awaiter(&ksched_waiter, __ksched_tick);
set_ksched_alarm();
+ corealloc_init();
+ spin_unlock(&sched_lock);
- /* Ghetto old idle core init */
- /* Init idle cores. Core 0 is the management core. */
- spin_lock(&idle_lock);
-#ifdef __CONFIG_DISABLE_SMT__
- /* assumes core0 is the only management core (NIC and monitor functionality
- * are run there too. it just adds the odd cores to the idlecoremap */
- assert(!(num_cpus % 2));
- // TODO: consider checking x86 for machines that actually hyperthread
- num_idlecores = num_cpus >> 1;
- #ifdef __CONFIG_ARSC_SERVER__
- // Dedicate one core (core 2) to sysserver, might be able to share wit NIC
- num_mgmtcores++;
- assert(num_cpus >= num_mgmtcores);
- send_kernel_message(2, (amr_t)arsc_server, 0,0,0, KMSG_ROUTINE);
- #endif
- for (int i = 0; i < num_idlecores; i++)
- idlecoremap[i] = (i * 2) + 1;
-#else
- // __CONFIG_DISABLE_SMT__
- #ifdef __CONFIG_NETWORKING__
- num_mgmtcores++; // Next core is dedicated to the NIC
- assert(num_cpus >= num_mgmtcores);
- #endif
- #ifdef __CONFIG_APPSERVER__
- #ifdef __CONFIG_DEDICATED_MONITOR__
- num_mgmtcores++; // Next core dedicated to running the kernel monitor
- assert(num_cpus >= num_mgmtcores);
- // Need to subtract 1 from the num_mgmtcores # to get the cores index
- send_kernel_message(num_mgmtcores-1, (amr_t)monitor, 0,0,0, KMSG_ROUTINE);
- #endif
- #endif
- #ifdef __CONFIG_ARSC_SERVER__
- // Dedicate one core (core 2) to sysserver, might be able to share with NIC
- num_mgmtcores++;
- assert(num_cpus >= num_mgmtcores);
- send_kernel_message(num_mgmtcores-1, (amr_t)arsc_server, 0,0,0, KMSG_ROUTINE);
- #endif
- num_idlecores = num_cpus - num_mgmtcores;
- for (int i = 0; i < num_idlecores; i++)
- idlecoremap[i] = i + num_mgmtcores;
-#endif /* __CONFIG_DISABLE_SMT__ */
- spin_unlock(&idle_lock);
- return;
+#ifdef CONFIG_ARSC_SERVER
+ /* Most likely we'll have a syscall and a process that dedicates itself to
+ * running this. Or if it's a kthread, we don't need a core. */
+ #error "Find a way to get a core. Probably a syscall to run a server."
+ int arsc_coreid = get_any_idle_core();
+ assert(arsc_coreid >= 0);
+ send_kernel_message(arsc_coreid, arsc_server, 0, 0, 0, KMSG_ROUTINE);
+ printk("Using core %d for the ARSC server\n", arsc_coreid);
+#endif /* CONFIG_ARSC_SERVER */
}
/* Round-robins on whatever list it's on */
static void add_to_list(struct proc *p, struct proc_list *new)
{
+ assert(!(p->ksched_data.cur_list));
TAILQ_INSERT_TAIL(new, p, ksched_data.proc_link);
p->ksched_data.cur_list = new;
}
{
assert(p->ksched_data.cur_list == old);
TAILQ_REMOVE(old, p, ksched_data.proc_link);
+ p->ksched_data.cur_list = 0;
}
static void switch_lists(struct proc *p, struct proc_list *old,
/* Removes from whatever list p is on */
static void remove_from_any_list(struct proc *p)
{
- assert(p->ksched_data.cur_list);
- TAILQ_REMOVE(p->ksched_data.cur_list, p, ksched_data.proc_link);
+ if (p->ksched_data.cur_list) {
+ TAILQ_REMOVE(p->ksched_data.cur_list, p, ksched_data.proc_link);
+ p->ksched_data.cur_list = 0;
+ }
}
-void register_proc(struct proc *p)
+/************** Process Management Callbacks **************/
+/* a couple notes:
+ * - the proc lock is NOT held for any of these calls. currently, there is no
+ * lock ordering between the sched lock and the proc lock. since the proc
+ * code doesn't know what we do, it doesn't hold its lock when calling our
+ * CBs.
+ * - since the proc lock isn't held, the proc could be dying, which means we
+ * will receive a __sched_proc_destroy() either before or after some of these
+ * other CBs. the CBs related to list management need to check and abort if
+ * DYING */
+void __sched_proc_register(struct proc *p)
{
+ assert(!proc_is_dying(p)); /* shouldn't be able to happen yet */
/* one ref for the proc's existence, cradle-to-grave */
proc_incref(p, 1); /* need at least this OR the 'one for existing' */
spin_lock(&sched_lock);
+ corealloc_proc_init(p);
add_to_list(p, &unrunnable_scps);
spin_unlock(&sched_lock);
}
-/* TODO: the proc lock is currently held for sched and register */
-/* sched_scp tells us to try and run the scp
- * TODO: change this horrible name */
-void schedule_scp(struct proc *p)
-{
- spin_lock(&sched_lock);
- printd("Scheduling PID: %d\n", p->pid);
- switch_lists(p, &unrunnable_scps, &runnable_scps);
- spin_unlock(&sched_lock);
-}
-
/* Returns 0 if it succeeded, an error code otherwise. */
-int proc_change_to_m(struct proc *p)
+void __sched_proc_change_to_m(struct proc *p)
{
- int retval;
spin_lock(&sched_lock);
- /* Should only be necessary to lock around the change_to_m call. It's
- * definitely necessary to hold the sched lock the whole time - need to
- * atomically change the proc's state and have the ksched take action (and
- * not squeeze a proc_destroy in there or something). */
- spin_lock(&p->proc_lock);
- retval = __proc_change_to_m(p);
- spin_unlock(&p->proc_lock);
- if (retval) {
- /* Failed for some reason. */
+ /* Need to make sure they aren't dying. if so, we already dealt with their
+ * list membership, etc (or soon will). taking advantage of the 'immutable
+ * state' of dying (so long as refs are held). */
+ if (proc_is_dying(p)) {
spin_unlock(&sched_lock);
- return retval;
+ return;
}
/* Catch user bugs */
if (!p->procdata->res_req[RES_CORES].amt_wanted) {
* probably a bug, at this stage in development, to do o/w. */
remove_from_list(p, &unrunnable_scps);
//remove_from_any_list(p); /* ^^ instead of this */
- add_to_list(p, &all_mcps);
+ add_to_list(p, primary_mcps);
spin_unlock(&sched_lock);
//poke_ksched(p, RES_CORES);
- return retval;
}
-/* Destroys the given process. This may be called from another process, a light
- * kernel thread (no real process context), asynchronously/cross-core, or from
- * the process on its own core.
+/* Sched callback called when the proc dies. pc_arr holds the cores the proc
+ * had, if any, and nr_cores tells us how many are in the array.
*
* An external, edible ref is passed in. when we return and they decref,
- * __proc_free will be called */
-void proc_destroy(struct proc *p)
+ * __proc_free will be called (when the last one is done). */
+void __sched_proc_destroy(struct proc *p, uint32_t *pc_arr, uint32_t nr_cores)
{
- uint32_t nr_cores_revoked = 0;
spin_lock(&sched_lock);
- spin_lock(&p->proc_lock);
- /* storage for pc_arr is alloced at decl, which is after grabbing the lock*/
- uint32_t pc_arr[p->procinfo->num_vcores];
- /* If this returns true, it means we successfully destroyed the proc */
- if (__proc_destroy(p, pc_arr, &nr_cores_revoked)) {
- /* Do our cleanup. note that proc_free won't run since we have an
- * external reference, passed in */
-
- /* Remove from whatever list we are on */
- remove_from_any_list(p);
- /* Drop the cradle-to-the-grave reference, jet-li */
- proc_decref(p);
- /* Put the cores back on the idlecore map. For future changes, be
- * careful with the idle_lock. It's safe to call this here or outside
- * the sched lock (for now). */
- if (nr_cores_revoked)
- put_idle_cores(pc_arr, nr_cores_revoked);
+ /* Unprovision any cores. Note this is different than track_core_dealloc.
+ * The latter does bookkeeping when an allocation changes. This is a
+ * bulk *provisioning* change. */
+ __unprovision_all_cores(p);
+ /* Remove from whatever list we are on (if any - might not be on one if it
+ * was in the middle of __run_mcp_sched) */
+ remove_from_any_list(p);
+ if (nr_cores)
+ __track_core_dealloc_bulk(p, pc_arr, nr_cores);
+ spin_unlock(&sched_lock);
+ /* Drop the cradle-to-the-grave reference, jet-li */
+ proc_decref(p);
+}
+
+/* ksched callbacks. p just woke up and is UNLOCKED. */
+void __sched_mcp_wakeup(struct proc *p)
+{
+ spin_lock(&sched_lock);
+ if (proc_is_dying(p)) {
+ spin_unlock(&sched_lock);
+ return;
}
- spin_unlock(&p->proc_lock);
+ /* could try and prioritize p somehow (move it to the front of the list). */
spin_unlock(&sched_lock);
+ /* note they could be dying at this point too. */
+ poke(&ksched_poker, p);
+}
+
+/* ksched callbacks. p just woke up and is UNLOCKED. */
+void __sched_scp_wakeup(struct proc *p)
+{
+ spin_lock(&sched_lock);
+ if (proc_is_dying(p)) {
+ spin_unlock(&sched_lock);
+ return;
+ }
+ /* might not be on a list if it is new. o/w, it should be unrunnable */
+ remove_from_any_list(p);
+ add_to_list(p, &runnable_scps);
+ spin_unlock(&sched_lock);
+ /* we could be on a CG core, and all the mgmt cores could be halted. if we
+ * don't tell one of them about the new proc, they will sleep until the
+ * timer tick goes off. */
+ if (!management_core()) {
+ /* TODO: pick a better core and only send if halted.
+ *
+ * ideally, we'd know if a specific mgmt core is sleeping and wake it
+ * up. o/w, we could interrupt an already-running mgmt core that won't
+ * get to our new proc anytime soon. also, by poking core 0, a
+ * different mgmt core could remain idle (and this process would sleep)
+ * until its tick goes off */
+ send_ipi(0, I_POKE_CORE);
+ }
+}
+
+/* Callback to return a core to the ksched, which tracks it as idle and
+ * deallocated from p. The proclock is held (__core_req depends on that).
+ *
+ * This also is a trigger, telling us we have more cores. We could/should make
+ * a scheduling decision (or at least plan to). */
+void __sched_put_idle_core(struct proc *p, uint32_t coreid)
+{
+ spin_lock(&sched_lock);
+ __track_core_dealloc(p, coreid);
+ spin_unlock(&sched_lock);
+}
+
+/* Callback, bulk interface for put_idle. The proclock is held for this. */
+void __sched_put_idle_cores(struct proc *p, uint32_t *pc_arr, uint32_t num)
+{
+ spin_lock(&sched_lock);
+ __track_core_dealloc_bulk(p, pc_arr, num);
+ spin_unlock(&sched_lock);
+ /* could trigger a sched decision here */
}
/* mgmt/LL cores should call this to schedule the calling core and give it to an
* calling. returns TRUE if it scheduled a proc. */
static bool __schedule_scp(void)
{
+ // TODO: sort out lock ordering (proc_run_s also locks)
struct proc *p;
uint32_t pcoreid = core_id();
struct per_cpu_info *pcpui = &per_cpu_info[pcoreid];
- int8_t state = 0;
/* if there are any runnables, run them here and put any currently running
* SCP on the tail of the runnable queue. */
if ((p = TAILQ_FIRST(&runnable_scps))) {
- /* protect owning proc, cur_tf, etc. note this nests with the
- * calls in proc_yield_s */
- disable_irqsave(&state);
/* someone is currently running, dequeue them */
if (pcpui->owning_proc) {
+ spin_lock(&pcpui->owning_proc->proc_lock);
+ /* process might be dying, with a KMSG to clean it up waiting on
+ * this core. can't do much, so we'll attempt to restart */
+ if (proc_is_dying(pcpui->owning_proc)) {
+ run_as_rkm(run_scheduler);
+ spin_unlock(&pcpui->owning_proc->proc_lock);
+ return FALSE;
+ }
printd("Descheduled %d in favor of %d\n", pcpui->owning_proc->pid,
p->pid);
- __proc_yield_s(pcpui->owning_proc, pcpui->cur_tf);
+ __proc_set_state(pcpui->owning_proc, PROC_RUNNABLE_S);
+ /* Saving FP state aggressively. Odds are, the SCP was hit by an
+ * IRQ and has a HW ctx, in which case we must save. */
+ __proc_save_fpu_s(pcpui->owning_proc);
+ __proc_save_context_s(pcpui->owning_proc);
+ vcore_account_offline(pcpui->owning_proc, 0);
+ __seq_start_write(&p->procinfo->coremap_seqctr);
+ __unmap_vcore(p, 0);
+ __seq_end_write(&p->procinfo->coremap_seqctr);
+ spin_unlock(&pcpui->owning_proc->proc_lock);
/* round-robin the SCPs (inserts at the end of the queue) */
switch_lists(pcpui->owning_proc, &unrunnable_scps, &runnable_scps);
clear_owning_proc(pcoreid);
* proc_run_s would pick it up. This way is a bit safer for
* future changes, but has an extra (empty) TLB flush. */
abandon_core();
- }
+ }
/* Run the new proc */
switch_lists(p, &runnable_scps, &unrunnable_scps);
printd("PID of the SCP i'm running: %d\n", p->pid);
proc_run_s(p); /* gives it core we're running on */
- enable_irqsave(&state);
return TRUE;
}
return FALSE;
}
-/* Something has changed, and for whatever reason the scheduler should
- * reevaluate things.
- *
- * Don't call this from interrupt context (grabs proclocks). */
-void schedule(void)
+/* Returns how many new cores p needs. This doesn't lock the proc, so your
+ * answer might be stale. */
+static uint32_t get_cores_needed(struct proc *p)
+{
+ uint32_t amt_wanted, amt_granted;
+ amt_wanted = p->procdata->res_req[RES_CORES].amt_wanted;
+ /* Help them out - if they ask for something impossible, give them 1 so they
+ * can make some progress. (this is racy, and unnecessary). */
+ if (amt_wanted > p->procinfo->max_vcores) {
+ printk("[kernel] proc %d wanted more than max, wanted %d\n", p->pid,
+ amt_wanted);
+ p->procdata->res_req[RES_CORES].amt_wanted = 1;
+ amt_wanted = 1;
+ }
+ /* There are a few cases where amt_wanted is 0, but they are still RUNNABLE
+ * (involving yields, events, and preemptions). In these cases, give them
+ * at least 1, so they can make progress and yield properly. If they are
+ * not WAITING, they did not yield and may have missed a message. */
+ if (!amt_wanted) {
+ /* could ++, but there could be a race and we don't want to give them
+ * more than they ever asked for (in case they haven't prepped) */
+ p->procdata->res_req[RES_CORES].amt_wanted = 1;
+ amt_wanted = 1;
+ }
+ /* amt_granted is racy - they could be *yielding*, but currently they can't
+ * be getting any new cores if the caller is in the mcp_ksched. this is
+ * okay - we won't accidentally give them more cores than they *ever* wanted
+ * (which could crash them), but our answer might be a little stale. */
+ amt_granted = p->procinfo->res_grant[RES_CORES];
+ /* Do not do an assert like this: it could fail (yield in progress): */
+ //assert(amt_granted == p->procinfo->num_vcores);
+ if (amt_wanted <= amt_granted)
+ return 0;
+ return amt_wanted - amt_granted;
+}
+
+/* Actual work of the MCP kscheduler. if we were called by poke_ksched, *arg
+ * might be the process who wanted special service. this would be the case if
+ * we weren't already running the ksched. Sort of a ghetto way to "post work",
+ * such that it's an optimization. */
+static void __run_mcp_ksched(void *arg)
{
struct proc *p, *temp;
+ uint32_t amt_needed;
+ struct proc_list *temp_mcp_list;
+ /* locking to protect the MCP lists' integrity and membership */
spin_lock(&sched_lock);
- /* trivially try to handle the needs of all our MCPS. smarter schedulers
- * would do something other than FCFS */
- TAILQ_FOREACH_SAFE(p, &all_mcps, ksched_data.proc_link, temp) {
- printd("Ksched has MCP %08p (%d)\n", p, p->pid);
- if (!num_idlecores)
+ /* 2-pass scheme: check each proc on the primary list (FCFS). if they need
+ * nothing, put them on the secondary list. if they need something, rip
+ * them off the list, service them, and if they are still not dying, put
+ * them on the secondary list. We cull the entire primary list, so that
+ * when we start from the beginning each time, we aren't repeatedly checking
+ * procs we looked at on previous waves.
+ *
+ * TODO: we could modify this such that procs that we failed to service move
+ * to yet another list or something. We can also move the WAITINGs to
+ * another list and have wakeup move them back, etc. */
+ while (!TAILQ_EMPTY(primary_mcps)) {
+ TAILQ_FOREACH_SAFE(p, primary_mcps, ksched_data.proc_link, temp) {
+ if (p->state == PROC_WAITING) { /* unlocked peek at the state */
+ switch_lists(p, primary_mcps, secondary_mcps);
+ continue;
+ }
+ amt_needed = get_cores_needed(p);
+ if (!amt_needed) {
+ switch_lists(p, primary_mcps, secondary_mcps);
+ continue;
+ }
+ /* o/w, we want to give cores to this proc */
+ remove_from_list(p, primary_mcps);
+ /* now it won't die, but it could get removed from lists and have
+ * its stuff unprov'd when we unlock */
+ proc_incref(p, 1);
+ /* GIANT WARNING: __core_req will unlock the sched lock for a bit.
+ * It will return with it locked still. We could unlock before we
+ * pass in, but they will relock right away. */
+ // notionally_unlock(&ksched_lock); /* for mouse-eyed viewers */
+ __core_request(p, amt_needed);
+ // notionally_lock(&ksched_lock);
+ /* Peeking at the state is okay, since we hold a ref. Once it is
+ * DYING, it'll remain DYING until we decref. And if there is a
+ * concurrent death, that will spin on the ksched lock (which we
+ * hold, and which protects the proc lists). */
+ if (!proc_is_dying(p))
+ add_to_list(p, secondary_mcps);
+ proc_decref(p); /* fyi, this may trigger __proc_free */
+ /* need to break: the proc lists may have changed when we unlocked
+ * in core_req in ways that the FOREACH_SAFE can't handle. */
break;
- /* TODO: might use amt_wanted as a proxy. right now, they have
- * amt_wanted == 1, even though they are waiting.
- * TODO: this is RACY too - just like with DYING. */
- if (p->state == PROC_WAITING)
- continue;
- __core_request(p);
+ }
}
- if (management_core())
- __schedule_scp();
+ /* at this point, we moved all the procs over to the secondary list, and
+ * attempted to service the ones that wanted something. now just swap the
+ * lists for the next invocation of the ksched. */
+ temp_mcp_list = primary_mcps;
+ primary_mcps = secondary_mcps;
+ secondary_mcps = temp_mcp_list;
spin_unlock(&sched_lock);
}
-/* A process is asking the ksched to look at its resource desires. The
- * scheduler is free to ignore this, for its own reasons, so long as it
- * eventually gets around to looking at resource desires. */
-void poke_ksched(struct proc *p, int res_type)
+/* Something has changed, and for whatever reason the scheduler should
+ * reevaluate things.
+ *
+ * Don't call this if you are processing a syscall or otherwise care about your
+ * kthread variables, cur_proc/owning_proc, etc.
+ *
+ * Don't call this from interrupt context (grabs proclocks). */
+void run_scheduler(void)
{
- /* TODO: probably want something to trigger all res_types */
- spin_lock(&sched_lock);
- switch (res_type) {
- case RES_CORES:
- /* ignore core requests from non-mcps (note we have races if we ever
- * allow procs to switch back). */
- if (!__proc_is_mcp(p))
- break;
- __core_request(p);
- break;
- default:
- break;
+ /* MCP scheduling: post work, then poke. for now, i just want the func to
+ * run again, so merely a poke is sufficient. */
+ poke(&ksched_poker, 0);
+ if (management_core()) {
+ spin_lock(&sched_lock);
+ __schedule_scp();
+ spin_unlock(&sched_lock);
}
- spin_unlock(&sched_lock);
}
-/* Proc p just woke up (due to an event). Our dumb ksched will just try to deal
- * with its core desires.
- * TODO: this may get called multiple times per unblock */
-void ksched_proc_unblocked(struct proc *p)
+/* A process is asking the ksched to look at its resource desires. The
+ * scheduler is free to ignore this, for its own reasons, so long as it
+ * eventually gets around to looking at resource desires. */
+void poke_ksched(struct proc *p, unsigned int res_type)
{
- /* TODO: this now gets called when an _S unblocks. schedule_scp() also gets
- * called, so the process is on the _S runqueue. Might merge the two in the
- * future. */
- poke_ksched(p, RES_CORES);
+ /* ignoring res_type for now. could post that if we wanted (would need some
+ * other structs/flags) */
+ if (!__proc_is_mcp(p))
+ return;
+ poke(&ksched_poker, p);
}
/* The calling cpu/core has nothing to do and plans to idle/halt. This is an
* the 'call of the giraffe' suffices. */
}
-/* Helper function to return a core to the idlemap. It causes some more lock
- * acquisitions (like in a for loop), but it's a little easier. Plus, one day
- * we might be able to do this without locks (for the putting).
- *
- * This is a trigger, telling us we have more cores. We could/should make a
- * scheduling decision (or at least plan to). */
-void put_idle_core(uint32_t coreid)
-{
- spin_lock(&idle_lock);
- idlecoremap[num_idlecores++] = coreid;
- spin_unlock(&idle_lock);
-}
-
-/* Helper for put_idle and core_req. */
-static void __put_idle_cores(uint32_t *pc_arr, uint32_t num)
-{
- spin_lock(&idle_lock);
- for (int i = 0; i < num; i++)
- idlecoremap[num_idlecores++] = pc_arr[i];
- spin_unlock(&idle_lock);
-}
-
-/* Bulk interface for put_idle */
-void put_idle_cores(uint32_t *pc_arr, uint32_t num)
-{
- /* could trigger a sched decision here */
- __put_idle_cores(pc_arr, num);
-}
-
/* Available resources changed (plus or minus). Some parts of the kernel may
* call this if a particular resource that is 'quantity-based' changes. Things
* like available RAM to processes, bandwidth, etc. Cores would probably be
printk("[kernel] ksched doesn't track any resources yet!\n");
}
-/* Normally it'll be the max number of CG cores ever */
-uint32_t max_vcores(struct proc *p)
-{
-#ifdef __CONFIG_DISABLE_SMT__
- return num_cpus >> 1;
-#else
- return MAX(1, num_cpus - num_mgmtcores);
-#endif /* __CONFIG_DISABLE_SMT__ */
-}
-
-/* Ghetto helper, just hands out up to 'amt_new' cores (no sense of locality or
- * anything) */
-static uint32_t get_idle_cores(struct proc *p, uint32_t *pc_arr,
- uint32_t amt_new)
-{
- uint32_t num_granted = 0;
- spin_lock(&idle_lock);
- for (int i = 0; i < num_idlecores && i < amt_new; i++) {
- /* grab the last one on the list */
- pc_arr[i] = idlecoremap[num_idlecores - 1];
- num_idlecores--;
- num_granted++;
- }
- spin_unlock(&idle_lock);
- return num_granted;
-}
-
-/* This deals with a request for more cores. The request is already stored in
- * the proc's amt_wanted (it is compared to amt_granted). */
-static void __core_request(struct proc *p)
+/* This deals with a request for more cores. The amt of new cores needed is
+ * passed in. The ksched lock is held, but we are free to unlock if we want
+ * (and we must, if calling out of the ksched to anything high-level).
+ *
+ * Side note: if we want to warn, then we can't deal with this proc's prov'd
+ * cores until we wait til the alarm goes off. would need to put all
+ * alarmed cores on a list and wait til the alarm goes off to do the full
+ * preempt. and when those cores come in voluntarily, we'd need to know to
+ * give them to this proc. */
+static void __core_request(struct proc *p, uint32_t amt_needed)
{
- uint32_t num_granted, amt_wanted, amt_granted;
- uint32_t corelist[num_cpus];
-
- /* TODO: consider copy-in for amt_wanted too. */
- amt_wanted = p->procdata->res_req[RES_CORES].amt_wanted;
- amt_granted = p->procinfo->res_grant[RES_CORES];
-
- /* Help them out - if they ask for something impossible, give them 1 so they
- * can make some progress. (this is racy). */
- if (amt_wanted > p->procinfo->max_vcores) {
- p->procdata->res_req[RES_CORES].amt_wanted = 1;
+ uint32_t nr_to_grant = 0;
+ uint32_t corelist[num_cores];
+ uint32_t pcoreid;
+ struct proc *proc_to_preempt;
+ bool success;
+ /* we come in holding the ksched lock, and we hold it here to protect
+ * allocations and provisioning. */
+ /* get all available cores from their prov_not_alloc list. the list might
+ * change when we unlock (new cores added to it, or the entire list emptied,
+ * but no core allocations will happen (we hold the poke)). */
+ while (nr_to_grant != amt_needed) {
+ /* Find the next best core to allocate to p. It may be a core
+ * provisioned to p, and it might not be. */
+ pcoreid = __find_best_core_to_alloc(p);
+ /* If no core is returned, we know that there are no more cores to give
+ * out, so we exit the loop. */
+ if (pcoreid == -1)
+ break;
+ /* If the pcore chosen currently has a proc allocated to it, we know
+ * it must be provisioned to p, but not allocated to it. We need to try
+ * to preempt. After this block, the core will be track_dealloc'd and
+ * on the idle list (regardless of whether we had to preempt or not) */
+ if (get_alloc_proc(pcoreid)) {
+ proc_to_preempt = get_alloc_proc(pcoreid);
+ /* would break both preemption and maybe the later decref */
+ assert(proc_to_preempt != p);
+ /* need to keep a valid, external ref when we unlock */
+ proc_incref(proc_to_preempt, 1);
+ spin_unlock(&sched_lock);
+ /* sending no warning time for now - just an immediate preempt. */
+ success = proc_preempt_core(proc_to_preempt, pcoreid, 0);
+ /* reaquire locks to protect provisioning and idle lists */
+ spin_lock(&sched_lock);
+ if (success) {
+ /* we preempted it before the proc could yield or die.
+ * alloc_proc should not have changed (it'll change in death and
+ * idle CBs). the core is not on the idle core list. (if we
+ * ever have proc alloc lists, it'll still be on the old proc's
+ * list). */
+ assert(get_alloc_proc(pcoreid));
+ /* regardless of whether or not it is still prov to p, we need
+ * to note its dealloc. we are doing some excessive checking of
+ * p == prov_proc, but using this helper is a lot clearer. */
+ __track_core_dealloc(proc_to_preempt, pcoreid);
+ } else {
+ /* the preempt failed, which should only happen if the pcore was
+ * unmapped (could be dying, could be yielding, but NOT
+ * preempted). whoever unmapped it also triggered (or will soon
+ * trigger) a track_core_dealloc and put it on the idle list.
+ * Our signal for this is get_alloc_proc() being 0. We need to
+ * spin and let whoever is trying to free the core grab the
+ * ksched lock. We could use an 'ignore_next_idle' flag per
+ * sched_pcore, but it's not critical anymore.
+ *
+ * Note, we're relying on us being the only preemptor - if the
+ * core was unmapped by *another* preemptor, there would be no
+ * way of knowing the core was made idle *yet* (the success
+ * branch in another thread). likewise, if there were another
+ * allocator, the pcore could have been put on the idle list and
+ * then quickly removed/allocated. */
+ cmb();
+ while (get_alloc_proc(pcoreid)) {
+ /* this loop should be very rare */
+ spin_unlock(&sched_lock);
+ udelay(1);
+ spin_lock(&sched_lock);
+ }
+ }
+ /* no longer need to keep p_to_pre alive */
+ proc_decref(proc_to_preempt);
+ /* might not be prov to p anymore (rare race). pcoreid is idle - we
+ * might get it later, or maybe we'll give it to its rightful proc*/
+ if (get_prov_proc(pcoreid) != p)
+ continue;
+ }
+ /* At this point, the pcore is idle, regardless of how we got here
+ * (successful preempt, failed preempt, or it was idle in the first
+ * place). We also know the core is still provisioned to us. Lets add
+ * it to the corelist for p (so we can give it to p in bulk later), and
+ * track its allocation with p (so our internal data structures stay in
+ * sync). We rely on the fact that we are the only allocator (pcoreid is
+ * still idle, despite (potentially) unlocking during the preempt
+ * attempt above). It is guaranteed to be track_dealloc'd()
+ * (regardless of how we got here). */
+ corelist[nr_to_grant] = pcoreid;
+ nr_to_grant++;
+ __track_core_alloc(p, pcoreid);
}
- /* if they are satisfied, we're done. There's a slight chance they have
- * cores, but they aren't running (sched gave them cores while they were
- * yielding, and now we see them on the run queue). */
- if (amt_wanted <= amt_granted)
- return;
- /* Otherwise, see what they want, and try to give out as many as possible.
- * Current models are simple - it's just a raw number of cores, and we just
- * give out what we can. */
- num_granted = get_idle_cores(p, corelist, amt_wanted - amt_granted);
/* Now, actually give them out */
- if (num_granted) {
+ if (nr_to_grant) {
+ /* Need to unlock before calling out to proc code. We are somewhat
+ * relying on being the only one allocating 'thread' here, since another
+ * allocator could have seen these cores (if they are prov to some proc)
+ * and could be trying to give them out (and assuming they are already
+ * on the idle list). */
+ spin_unlock(&sched_lock);
/* give them the cores. this will start up the extras if RUNNING_M. */
spin_lock(&p->proc_lock);
/* if they fail, it is because they are WAITING or DYING. we could give
* ksched, we'll just put them back on the pile and return. Note, the
* ksched could check the states after locking, but it isn't necessary:
* just need to check at some point in the ksched loop. */
- if (__proc_give_cores(p, corelist, num_granted)) {
- __put_idle_cores(corelist, num_granted);
+ if (__proc_give_cores(p, corelist, nr_to_grant)) {
+ spin_unlock(&p->proc_lock);
+ /* we failed, put the cores and track their dealloc. lock is
+ * protecting those structures. */
+ spin_lock(&sched_lock);
+ __track_core_dealloc_bulk(p, corelist, nr_to_grant);
} else {
/* at some point after giving cores, call proc_run_m() (harmless on
* RUNNING_Ms). You can give small groups of cores, then run them
* (which is more efficient than interleaving runs with the gives
* for bulk preempted processes). */
__proc_run_m(p);
+ spin_unlock(&p->proc_lock);
+ /* main mcp_ksched wants this held (it came to __core_req held) */
+ spin_lock(&sched_lock);
}
- spin_unlock(&p->proc_lock);
}
+ /* note the ksched lock is still held */
+}
+
+/* Provision a core to a process. This function wraps the primary logic
+ * implemented in __provision_core, with a lock, error checking, etc. */
+int provision_core(struct proc *p, uint32_t pcoreid)
+{
+ /* Make sure we aren't asking for something that doesn't exist (bounds check
+ * on the pcore array) */
+ if (!(pcoreid < num_cores)) {
+ set_errno(ENXIO);
+ return -1;
+ }
+ /* Don't allow the provisioning of LL cores */
+ if (is_ll_core(pcoreid)) {
+ set_errno(EBUSY);
+ return -1;
+ }
+ /* Note the sched lock protects the tailqs for all procs in this code.
+ * If we need a finer grained sched lock, this is one place where we could
+ * have a different lock */
+ spin_lock(&sched_lock);
+ __provision_core(p, pcoreid);
+ spin_unlock(&sched_lock);
+ return 0;
}
/************** Debugging **************/
printk("Runnable _S PID: %d\n", p->pid);
TAILQ_FOREACH(p, &unrunnable_scps, ksched_data.proc_link)
printk("Unrunnable _S PID: %d\n", p->pid);
- TAILQ_FOREACH(p, &all_mcps, ksched_data.proc_link)
- printk("MCP PID: %d\n", p->pid);
+ TAILQ_FOREACH(p, primary_mcps, ksched_data.proc_link)
+ printk("Primary MCP PID: %d\n", p->pid);
+ TAILQ_FOREACH(p, secondary_mcps, ksched_data.proc_link)
+ printk("Secondary MCP PID: %d\n", p->pid);
spin_unlock(&sched_lock);
return;
}
-void print_idlecoremap(void)
-{
- spin_lock(&idle_lock);
- printk("There are %d idle cores.\n", num_idlecores);
- for (int i = 0; i < num_idlecores; i++)
- printk("idlecoremap[%d] = %d\n", i, idlecoremap[i]);
- spin_unlock(&idle_lock);
-}
-
void print_resources(struct proc *p)
{
printk("--------------------\n");
void print_all_resources(void)
{
/* Hash helper */
- void __print_resources(void *item)
+ void __print_resources(void *item, void *opaque)
{
print_resources((struct proc*)item);
}
spin_lock(&pid_hash_lock);
- hash_for_each(pid_hash, __print_resources);
+ hash_for_each(pid_hash, __print_resources, NULL);
spin_unlock(&pid_hash_lock);
}
+
+void next_core_to_alloc(uint32_t pcoreid)
+{
+ spin_lock(&sched_lock);
+ __next_core_to_alloc(pcoreid);
+ spin_unlock(&sched_lock);
+}
+
+void sort_idle_cores(void)
+{
+ spin_lock(&sched_lock);
+ __sort_idle_cores();
+ spin_unlock(&sched_lock);
+}
projects / akaros.git / blobdiff